Extensive-stage small cell lung cancer is an aggressive disease. Platinum-based chemotherapy has been the standard treatment for decades, but most patients experience disease progression within just six months and have few options after relapse. Extensive-stage small cell lung cancer (ES-SCLC) describes disease that has spread beyond the chest and cannot be contained within a single radiation field.
Immunotherapy offers hope but works for only some patients. Landmark clinical trials including IMpower133 and CASPIAN demonstrated that adding immune checkpoint inhibitors (ICIs) to standard chemotherapy prolonged overall survival in ES-SCLC patients. Multiple PD-1/PD-L1 inhibitors have since been approved, yet the overall survival benefit remains modest compared to what immunotherapy achieves in other lung cancer types.
No reliable biomarker exists to predict who will respond. In other cancers, markers like PD-L1 expression and tumor mutation burden (TMB) help identify immunotherapy responders, but neither has demonstrated predictive value specifically for ES-SCLC. This leaves clinicians without an accurate, non-invasive tool to identify which patients will actually benefit from adding immunotherapy.
CT-based radiomics as a possible solution. Radiomics - the extraction of large numbers of quantitative features from routine CT scan images - has shown promise in lung cancer research for tasks ranging from early detection to prognosis. This study explores whether combining radiomics features from CT images with standard clinical data can predict immunotherapy response more accurately than either source alone.
A multicenter retrospective study. Researchers at Shandong Cancer Hospital and Institute screened 449 patients with histologically confirmed ES-SCLC who received first-line chemoimmunotherapy between March 2020 and February 2024. After applying strict inclusion and exclusion criteria - including the requirement for accessible CT scans at baseline and after 2-3 treatment cycles - 159 patients were enrolled.
Training and external validation cohorts. The 159 patients were divided into a training set of 119 patients from the primary institution and an external test set of 40 patients from another center. Using an independent external test cohort is a critical strength: it verifies that the model generalizes beyond the population it was trained on, rather than just memorizing patterns from training data.
Defining treatment success. Rather than using short-term tumor response measures, the researchers used two-year overall survival as their primary outcome measure. This choice reflects evidence from clinical trials showing that immunotherapy in ES-SCLC tends to produce a 'long tail effect' - a subset of patients achieve durable long-term survival while others do not. The two-year threshold captures this distinction more meaningfully than progression-free survival alone.
Baseline patient characteristics. The median age was approximately 61 years, over 80% of patients were male, and roughly half had a history of smoking. All had stage IV disease at enrollment. Common metastatic sites included the liver (35-45%), bone (29-33%), and brain (26-28%), reflecting the widespread disease typical of ES-SCLC.
What radiomics measures. Radiomics extracts hundreds of quantitative features from medical images that are invisible to the human eye - including texture patterns, shape properties, and statistical distributions of pixel intensities. These features can reflect tumor biology such as cellularity, heterogeneity, and vascular structure at a level of detail beyond visual inspection.
Plain scan and venous phase CT images. Features were extracted from two types of CT images: plain scans (without contrast) and venous phase scans (after contrast injection, which highlights vascular structures). Tumor boundaries were carefully traced slice by slice by two experienced radiologists, with a senior radiologist resolving any disagreements - a process designed to ensure consistent and reproducible measurements.
Delta radiomics: capturing change over time. In addition to baseline features, the team computed 'delta radiomics' - features that quantify how tumor characteristics changed between baseline and post-treatment CT scans. Both absolute changes (numerical differences) and relative changes (proportional differences normalized to baseline) were calculated. This longitudinal approach captures tumor dynamics that static baseline measurements cannot.
Rigorous feature selection. Starting from 6,724 radiomic features total, a four-step selection process narrowed this to a manageable set. Features were first filtered for reproducibility (requiring an intraclass correlation coefficient above 0.8), then for variance, then for statistical association with outcomes, and finally refined using LASSO regularization with cross-validation. This systematic approach guards against overfitting and selects only the most informative features.
Four separate radiomic signatures. The researchers built and tested four radiomic models based on: features before treatment, features after treatment, absolute delta features (raw change), and relative delta features (proportional change). Each model's performance was measured using the area under the ROC curve (AUC) - a value from 0.5 (no better than chance) to 1.0 (perfect prediction).
Identifying clinical predictors. Alongside radiomics, standard clinical variables were analyzed using logistic regression. From among age, tumor markers, tumor size, and staging variables, two emerged as independent predictors: patient age and lymph node stage. These two clinical factors were carried forward into the combined model alongside the best-performing radiomics signatures.
Building the nomogram. The final combined model integrated two radiomic signatures (intratumoral region features before treatment, and relative delta radiomics) with two clinical factors (age and lymph node stage) into a nomogram - a graphical tool that assigns point values to each variable and generates an overall score that translates directly into predicted probability of a favorable immunotherapy response.
Validation and performance assessment. Model performance was evaluated using ROC curve analysis for discrimination, calibration curves for agreement between predicted and observed outcomes, and decision curve analysis (DCA) to assess clinical utility - that is, whether acting on the model's predictions would actually benefit patients at different decision thresholds.
Radiomics alone performs well but unevenly. The best single radiomic model - combining before-treatment and relative delta signatures - achieved an AUC of 0.908 in the training set and 0.807 in the external test set. This already outperforms the clinical-only model. However, the before-treatment model alone showed poor external test performance (AUC 0.476), highlighting the risk of relying on baseline features without longitudinal information.
Clinical model has limited accuracy. Using only age and lymph node stage, the clinical model achieved AUCs of 0.799 (training) and 0.693 (external test). This is meaningful but insufficient on its own, confirming that standard clinical variables capture only part of the picture of who will respond to immunotherapy.
The combined model outperforms both. The nomogram combining clinical and radiomic features achieved AUCs of 0.919 in training and 0.839 in the external test cohort - representing improvements of 0.120 and 0.146 over the clinical model alone, and 0.011 and 0.032 over the radiomics-only model. In the test cohort, the combined model also showed superior sensitivity (85.7%), specificity (66.7%), and overall accuracy (80.0%).
Survival differences confirmed by Kaplan-Meier analysis. Patients stratified as high-risk by the combined model score had substantially shorter overall survival than those classified as low-risk, and this separation was statistically significant in both the training and test cohorts. This confirms that the model captures clinically meaningful differences in immunotherapy benefit across patient subgroups.
Why investigate the immune microenvironment? Immune checkpoint inhibitors work by reactivating T cells that have been exhausted or suppressed within the tumor. The composition of the tumor immune microenvironment (TIME) - which immune cells are present and where they are located - fundamentally determines whether this reactivation can occur. The researchers sought to understand whether their model's predictions reflected underlying immune biology.
Tertiary lymphoid structures were investigated. Tertiary lymphoid structures (TLS) are organized immune cell aggregates that can form within tumors and have been linked to improved immunotherapy responses in some cancers. Multiplex immunohistochemistry was performed on baseline tissue samples from 33 patients to characterize TLS and immune cell populations within different tumor compartments.
Limited but suggestive immune correlations. No statistically significant associations were found between TLS presence and overall survival or combined model scores in this cohort. However, higher levels of CD23+ immune cells in tumor regions were associated with longer overall survival, and stromal CD8+ T cell infiltration showed a trend toward association with the model's risk classification. These findings are preliminary given the small sample size.
Why correlations were weak. Small cell lung cancer is characterized by an immunologically 'cold' tumor microenvironment - meaning it tends to have low immune cell infiltration compared to other cancers. Moreover, SCLC exhibits high intratumoral diversity and spatial heterogeneity, making it difficult to capture immune biology from limited biopsy samples. The limited sample size available for immune analysis further constrained statistical power.
Non-invasive and comprehensive. Most existing approaches to predicting immunotherapy response in SCLC require invasive tissue biopsies, which are subject to sampling error, limited tissue availability, and practical constraints in sick patients. CT imaging, by contrast, is routinely performed in all lung cancer patients at diagnosis and during follow-up, making it accessible, repeatable, and non-invasive.
CT captures whole-tumor information. Unlike a biopsy that samples only a small fragment of the tumor, CT imaging encompasses the entire lesion and surrounding tissue, capturing features related to tumor shape, internal heterogeneity, vascular responses, and peritumoral changes. This makes it inherently better suited for characterizing tumors that are biologically heterogeneous, as SCLC often is.
Why relative delta radiomics outperforms absolute measures. After 2-3 cycles of chemoimmunotherapy, SCLC tumors typically shrink rapidly due to chemosensitivity. This rapid regression makes post-treatment absolute measurements unstable and difficult to extract reliably. Relative delta features, which capture proportional changes normalized to baseline values, proved more robust because they adjust for this tumor-size variability.
Personalized treatment decision-making. The nomogram generates an individualized probability score for each patient, enabling oncologists to identify those most likely to achieve long-term benefit from immunotherapy versus those who may need alternative strategies or intensified monitoring. This kind of personalized risk stratification is increasingly important as more immunotherapy options become available for ES-SCLC.
Study limitations to acknowledge. As a retrospective study, the analysis is subject to potential selection bias and incomplete data. The final cohort of 159 patients is relatively small, which may limit the generalizability of the model and its statistical power to detect associations with immune microenvironment features. Variation in treatment regimens across patients - different PD-1/PD-L1 inhibitors were used - adds further heterogeneity.
Incomplete clinical variables. Progression-free survival and post-treatment laboratory parameters were not included in the analysis. Ki67 proliferation index and other pathological markers that might refine prognostic stratification were also excluded to maintain model simplicity. These omissions may have left some predictive information unexplored.
Future work planned. The authors plan to expand the patient cohort, collect more biological samples for immune analysis, and integrate multi-dimensional data sources to improve model comprehensiveness. Larger prospective studies with standardized CT imaging protocols across institutions will be needed to confirm the model's generalizability before clinical deployment.
A meaningful step forward. Despite its limitations, this study demonstrates for the first time that combining CT-based radiomics - particularly relative delta features that capture treatment-induced tumor changes - with basic clinical variables produces a practically useful model for predicting immunotherapy benefit in ES-SCLC. This framework could ultimately guide first-line treatment decisions and help reserve immunotherapy for those most likely to respond.